diff --git a/README.md b/README.md index 8bedf4a..198d121 100644 --- a/README.md +++ b/README.md @@ -8,7 +8,7 @@ # nonebot-plugin-llmchat -_✨ 支持多API预设、MCP协议、联网搜索的AI群聊插件 ✨_ +_✨ 支持多API预设、MCP协议、联网搜索、视觉模型的AI群聊插件 ✨_ @@ -17,38 +17,43 @@ _✨ 支持多API预设、MCP协议、联网搜索的AI群聊插件 ✨_ pypi -python +python +Ask DeepWiki ## 📖 介绍 +1. **支持LLM回复图片** + - 支持最新 Gemini 2.5 Flash Image (Nano Banana) 的图片回复 + - 支持图片上下文修改 + 1. **支持MCP协议** - 可以连接各种支持MCP协议的LLM工具 - 通过连接一些搜索MCP服务器可以实现在线搜索 - 兼容 Claude.app 的配置格式 -2. **多API预设支持** +1. **多API预设支持** - 可配置多个LLM服务预设（如不同模型/API密钥） - 支持运行时通过`API预设`命令热切换API配置 - 内置服务开关功能（预设名为`off`时停用） -3. **多种回复触发方式** +1. **多种回复触发方式** - @触发 + 随机概率触发 - 支持处理回复消息 - 群聊消息顺序处理，防止消息错乱 -4. **分群聊上下文记忆管理** +1. **分群聊上下文记忆管理** - 分群聊保留对话历史记录（可配置保留条数） - 自动合并未处理消息，降低API用量 - 支持`记忆清除`命令手动重置对话上下文 -5. **分段回复支持** +1. **分段回复支持** - 支持多段式回复（由LLM决定如何回复） - 可@群成员（由LLM插入） - 可选输出AI的思维过程（需模型支持） -6. **可自定义性格** +1. **可自定义性格** - 可动态修改群组专属系统提示词（`/修改设定`） - 支持自定义默认提示词 @@ -107,6 +112,8 @@ _✨ 支持多API预设、MCP协议、联网搜索的AI群聊插件 ✨_ | LLMCHAT__DEFAULT_PRESET | 否 | off | 默认使用的预设名称，配置为off则为关闭 | | LLMCHAT__RANDOM_TRIGGER_PROB | 否 | 0.05 | 默认随机触发概率 [0, 1] | | LLMCHAT__DEFAULT_PROMPT | 否 | 你的回答应该尽量简洁、幽默、可以使用一些语气词、颜文字。你应该拒绝回答任何政治相关的问题。 | 默认提示词 | +| LLMCHAT__BLACKLIST_USER_IDS | 否 | [] | 黑名单用户ID列表，机器人将不会处理黑名单用户的消息 | +| LLMCHAT__IGNORE_PREFIXES | 否 | [] | 需要忽略的消息前缀列表，匹配到这些前缀的消息不会处理 | | LLMCHAT__MCP_SERVERS | 否 | {} | MCP服务器配置，具体见下表 | 其中LLMCHAT__API_PRESETS为一个列表，每项配置有以下的配置项 @@ -119,6 +126,8 @@ _✨ 支持多API预设、MCP协议、联网搜索的AI群聊插件 ✨_ | max_tokens | 否 | 2048 | 最大响应token数 | | temperature | 否 | 0.7 | 生成温度 | | proxy | 否 | 无 | 请求API时使用的HTTP代理 | +| support_mcp | 否 | False | 是否支持MCP协议 | +| support_image | 否 | False | 是否支持图片输入 | LLMCHAT__MCP_SERVERS同样为一个dict，key为服务器名称，value配置的格式基本兼容 Claude.app 的配置格式，具体支持如下 @@ -151,10 +160,18 @@ LLMCHAT__MCP_SERVERS同样为一个dict，key为服务器名称，value配置的 "proxy": "http://10.0.0.183:7890" }, { - "name": "deepseek-r1", + "name": "deepseek-v1", "api_key": "sk-your-api-key", - "model_name": "deepseek-reasoner", - "api_base": "https://api.deepseek.com" + "model_name": "deepseek-chat", + "api_base": "https://api.deepseek.com", + "support_mcp": true + }, + { + "name": "some-vison-model", + "api_key": "sk-your-api-key", + "model_name": "some-vison-model", + "api_base": "https://some-vison-model.com/api", + "support_image": true } ] LLMCHAT__MCP_SERVERS=' @@ -195,4 +212,4 @@ LLMCHAT__MCP_SERVERS同样为一个dict，key为服务器名称，value配置的 ### 效果图 ![](img/mcp_demo.jpg) -![](img/demo.png) \ No newline at end of file +![](img/demo.png) diff --git a/nonebot_plugin_llmchat/__init__.py b/nonebot_plugin_llmchat/__init__.py old mode 100644 new mode 100755 index 879953a..e89e485 --- a/nonebot_plugin_llmchat/__init__.py +++ b/nonebot_plugin_llmchat/__init__.py @@ -1,10 +1,12 @@ import asyncio +import base64 from collections import defaultdict, deque from datetime import datetime import json import os import random import re +import ssl import time from typing import TYPE_CHECKING @@ -19,7 +21,7 @@ from nonebot import ( on_message, require, ) -from nonebot.adapters.onebot.v11 import GroupMessageEvent, Message +from nonebot.adapters.onebot.v11 import GroupMessageEvent, Message, MessageSegment from nonebot.adapters.onebot.v11.permission import GROUP_ADMIN, GROUP_OWNER from nonebot.params import CommandArg from nonebot.permission import SUPERUSER @@ -37,11 +39,14 @@ require("nonebot_plugin_apscheduler") from nonebot_plugin_apscheduler import scheduler if TYPE_CHECKING: - from openai.types.chat import ChatCompletionMessageParam + from openai.types.chat import ( + ChatCompletionContentPartParam, + ChatCompletionMessageParam, + ) __plugin_meta__ = PluginMetadata( name="llmchat", - description="支持多API预设、MCP协议、联网搜索的AI群聊插件", + description="支持多API预设、MCP协议、联网搜索、视觉模型、Nano Banana（生图模型）的AI群聊插件", usage="""@机器人 + 消息 开启对话""", type="application", homepage="https://github.com/FuQuan233/nonebot-plugin-llmchat", @@ -160,6 +165,16 @@ async def is_triggered(event: GroupMessageEvent) -> bool: if state.preset_name == "off": return False + # 黑名单用户 + if event.user_id in plugin_config.blacklist_user_ids: + return False + + # 忽略特定前缀的消息 + msg_text = event.get_plaintext().strip() + for prefix in plugin_config.ignore_prefixes: + if msg_text.startswith(prefix): + return False + state.past_events.append(event) # 原有@触发条件 @@ -176,7 +191,7 @@ async def is_triggered(event: GroupMessageEvent) -> bool: # 消息处理器 handler = on_message( rule=Rule(is_triggered), - priority=10, + priority=99, block=False, ) @@ -197,6 +212,46 @@ async def handle_message(event: GroupMessageEvent): task.add_done_callback(tasks.discard) tasks.add(task) +async def process_images(event: GroupMessageEvent) -> list[str]: + base64_images = [] + for segement in event.get_message(): + if segement.type == "image": + image_url = segement.data.get("url") or segement.data.get("file") + if image_url: + try: + # 处理高版本 httpx 的 [SSL: SSLV3_ALERT_HANDSHAKE_FAILURE] 报错 + ssl_context = ssl.create_default_context() + ssl_context.check_hostname = False + ssl_context.verify_mode = ssl.CERT_NONE + ssl_context.set_ciphers("DEFAULT@SECLEVEL=2") + + # 下载图片并将图片转换为base64 + async with httpx.AsyncClient(verify=ssl_context) as client: + response = await client.get(image_url, timeout=10.0) + if response.status_code != 200: + logger.error(f"下载图片失败: {image_url}, 状态码: {response.status_code}") + continue + image_data = response.content + base64_data = base64.b64encode(image_data).decode("utf-8") + base64_images.append(base64_data) + except Exception as e: + logger.error(f"处理图片时出错: {e}") + logger.debug(f"共处理 {len(base64_images)} 张图片") + return base64_images + +async def send_split_messages(message_handler, content: str): + """ + 将消息按分隔符分段并发送 + """ + logger.info(f"准备发送分段消息，分段数：{len(content.split(''))}") + for segment in content.split(""): + # 跳过空消息 + if not segment.strip(): + continue + segment = segment.strip() # 删除前后多余的换行和空格 + await asyncio.sleep(2) # 避免发送过快 + logger.debug(f"发送消息分段 内容：{segment[:50]}...") # 只记录前50个字符避免日志过大 + await message_handler.send(Message(segment)) async def process_messages(group_id: int): state = group_states[group_id] @@ -252,14 +307,28 @@ async def process_messages(group_id: int): {"role": "system", "content": systemPrompt} ] + while len(state.history) > 0 and state.history[0]["role"] != "user": + state.history.popleft() + messages += list(state.history)[-plugin_config.history_size * 2 :] # 没有未处理的消息说明已经被处理了，跳过 if state.past_events.__len__() < 1: break + content: list[ChatCompletionContentPartParam] = [] + # 将机器人错过的消息推送给LLM - content = ",".join([format_message(ev) for ev in state.past_events]) + past_events_snapshot = list(state.past_events) + for ev in past_events_snapshot: + text_content = format_message(ev) + content.append({"type": "text", "text": text_content}) + + # 将消息中的图片转成 base64 + if preset.support_image: + base64_images = await process_images(ev) + for base64_image in base64_images: + content.append({"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{base64_image}"}}) new_messages: list[ChatCompletionMessageParam] = [ {"role": "user", "content": content} @@ -301,7 +370,7 @@ async def process_messages(group_id: int): # 发送LLM调用工具时的回复，一般没有 if message.content: - await handler.send(Message(message.content)) + await send_split_messages(handler, message.content) # 处理每个工具调用 for tool_call in message.tool_calls: @@ -317,7 +386,7 @@ async def process_messages(group_id: int): new_messages.append({ "role": "tool", "tool_call_id": tool_call.id, - "content": str(result.content) + "content": str(result) }) # 将工具调用的结果交给 LLM @@ -338,10 +407,18 @@ async def process_messages(group_id: int): or matched_reasoning_content ) - new_messages.append({ + llm_reply: ChatCompletionMessageParam = { "role": "assistant", "content": reply, - }) + } + + reply_images = getattr(response.choices[0].message, "images", None) + + if reply_images: + # openai的sdk里的assistant消息暂时没有images字段，需要单独处理 + llm_reply["images"] = reply_images # pyright: ignore[reportGeneralTypeIssues] + + new_messages.append(llm_reply) # 请求成功后再保存历史记录，保证user和assistant穿插，防止R1模型报错 for message in new_messages: @@ -361,20 +438,15 @@ async def process_messages(group_id: int): logger.error(f"合并转发消息发送失败：\n{e!s}\n") assert reply is not None - logger.info( - f"准备发送回复消息 群号：{group_id} 消息分段数：{len(reply.split(''))}" - ) - for r in reply.split(""): - # 似乎会有空消息的情况导致string index out of range异常 - if len(r) == 0 or r.isspace(): - continue - # 删除前后多余的换行和空格 - r = r.strip() - await asyncio.sleep(2) - logger.debug( - f"发送消息分段 内容：{r[:50]}..." - ) # 只记录前50个字符避免日志过大 - await handler.send(Message(r)) + await send_split_messages(handler, reply) + + if reply_images: + logger.debug(f"API响应 图片数：{len(reply_images)}") + for i, image in enumerate(reply_images, start=1): + logger.debug(f"正在发送第{i}张图片") + image_base64 = image["image_url"]["url"].removeprefix("data:image/png;base64,") + image_msg = MessageSegment.image(base64.b64decode(image_base64)) + await handler.send(image_msg) except Exception as e: logger.opt(exception=e).error(f"API请求失败 群号：{group_id}") @@ -411,7 +483,7 @@ async def handle_preset(event: GroupMessageEvent, args: Message = CommandArg()): edit_preset_handler = on_command( "修改设定", - priority=99, + priority=1, block=True, permission=(SUPERUSER | GROUP_ADMIN | GROUP_OWNER), ) @@ -428,7 +500,7 @@ async def handle_edit_preset(event: GroupMessageEvent, args: Message = CommandAr reset_handler = on_command( "记忆清除", - priority=99, + priority=1, block=True, permission=(SUPERUSER | GROUP_ADMIN | GROUP_OWNER), ) @@ -445,7 +517,7 @@ async def handle_reset(event: GroupMessageEvent, args: Message = CommandArg()): set_prob_handler = on_command( "设置主动回复概率", - priority=99, + priority=1, block=True, permission=(SUPERUSER | GROUP_ADMIN | GROUP_OWNER), ) diff --git a/nonebot_plugin_llmchat/config.py b/nonebot_plugin_llmchat/config.py old mode 100644 new mode 100755 index c802a8d..d658875 --- a/nonebot_plugin_llmchat/config.py +++ b/nonebot_plugin_llmchat/config.py @@ -12,6 +12,7 @@ class PresetConfig(BaseModel): temperature: float = Field(0.7, description="生成温度（0-2]") proxy: str = Field("", description="HTTP代理服务器") support_mcp: bool = Field(False, description="是否支持MCP") + support_image: bool = Field(False, description="是否支持图片输入") class MCPServerConfig(BaseModel): """MCP服务器配置""" @@ -42,6 +43,11 @@ class ScopedConfig(BaseModel): description="默认提示词", ) mcp_servers: dict[str, MCPServerConfig] = Field({}, description="MCP服务器配置") + blacklist_user_ids: set[int] = Field(set(), description="黑名单用户ID列表") + ignore_prefixes: list[str] = Field( + default_factory=list, + description="需要忽略的消息前缀列表，匹配到这些前缀的消息不会处理" + ) class Config(BaseModel): diff --git a/nonebot_plugin_llmchat/mcpclient.py b/nonebot_plugin_llmchat/mcpclient.py index 7031d34..55e1b44 100644 --- a/nonebot_plugin_llmchat/mcpclient.py +++ b/nonebot_plugin_llmchat/mcpclient.py @@ -1,3 +1,4 @@ +import asyncio from contextlib import AsyncExitStack from mcp import ClientSession, StdioServerParameters @@ -64,9 +65,13 @@ class MCPClient: server_name, real_tool_name = tool_name.split("___") logger.info(f"正在服务器[{server_name}]上调用工具[{real_tool_name}]") session = self.sessions[server_name] - response = await session.call_tool(real_tool_name, tool_args) + try: + response = await asyncio.wait_for(session.call_tool(real_tool_name, tool_args), timeout=30) + except asyncio.TimeoutError: + logger.error(f"调用工具[{real_tool_name}]超时") + return f"调用工具[{real_tool_name}]超时" logger.debug(f"工具[{real_tool_name}]调用完成，响应: {response}") - return response + return response.content def get_friendly_name(self, tool_name: str): server_name, real_tool_name = tool_name.split("___") diff --git a/pyproject.toml b/pyproject.toml index 4875a8f..bb49e2f 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [tool.poetry] name = "nonebot-plugin-llmchat" -version = "0.2.1" +version = "0.3.0" description = "Nonebot AI group chat plugin supporting multiple API preset configurations" license = "GPL" authors = ["FuQuan i@fuquan.moe"]